Method of segmenting an audio stream

• US 20030171936A1
• Filed: 02/21/2003
• Published: 09/11/2003
• Est. Priority Date: 02/21/2002
• Status: Active Grant

• Current Assignee: LG Electronics, Inc. (LG Corporation)
• Original Assignee: LG Electronics, Inc. (LG Corporation)
• Inventors: Tikhotsky, Anatoli I., Sall, Mikhael A., Viktorov, Andrei B., Maiboroda, Alexandr L., Redkov, Victor V., Gramnitskiy, Sergei N.
• Granted Patent: US 7,346,516 B2
• Time in Patent Office: Days
• US Class Current: 704/500
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 25/48 specially adapted for particular use

Abstract

Disclosed herein is a segmentation method, which divides an input audio stream into segments containing different homogeneous signals. The main objective of this method is localization of segments with stationary properties. This method seeks all no-stationary points or intervals in the audio stream and creates a list of segments. The obtained list of segments can be used as an input data for the following procedures, such as classification, speech/music/noise attribution and so on. The proposed segmentation method is based on the analysis of audio signal statistical features variation and comprises three main stages: stage of first-grade characteristics calculation, stage of second-grade characteristics calculation and stage of decision-making.

Claims

1. A method of segmentation of an audio stream, wherein the segmentation is the division of the audio stream into segments containing different homogeneous signals.

2. The method according to claim 1, wherein the audio stream is a sequence of digital samples which are broadcasted or recorded using some media.

3. The method according to claim 1, wherein the audio stream segmentation is performed in three stages:

the stage of the first-grade characteristic calculation, the stage of the second-grade characteristic calculation, and the stage of the decision-making analysis.

4. The method according to claim 3, wherein the stage of the first-grade characteristic calculation is performed by the division of the audio stream into frames for which of them the audio feature vector is calculated.

5. The method according to claim 4, wherein the audio feature vector consists of five formant frequencies, the first and the second reflection coefficients, the energy of the prediction error coefficient, and the preemphasized energy ratio coefficient.

6. The method according to claim 3, wherein the stage of the second-grade characteristic calculation is performed in the sequence of the predefined and not overlapped windows, each of them consists from definite number of said frames with said audio feature vectors calculated at the stage of the first-grade characteristic calculation.

7. The method according to claim 6, wherein the stage of the second-grade characteristic calculation consists in the calculation of the statistical feature vector for each said window.

8. The method according to claim 7, wherein the statistical feature vector consists from two sub-vectors, firsts of which consists from mean values of the formant frequencies, and dispersions of the formant frequencies, and the second of said sub-vectors consists from difference between maximal and minimal values of the second reflection coefficient multiplied by the mean value of the second reflection coefficient, product of the mean value and the dispersion of the energy of the prediction error coefficient, sum of the modules of differences between said energies of the prediction error coefficients for said neighboring frames divided by the sum of the modules of said energies of the prediction error coefficients, difference between maximal and minimal values of said preemphasized energy ratio coefficients, and number of said frames in the window in which the first reflection coefficients outnumber predefined a predefined threshold value.

9. The method according to claim 3, wherein the stage of the decision-making analysis is performed in three sub-stages:

the sub-stage of initial segmentation, the sub-stage of accurate segmentation, and the sub-stage of the internal markers definition.

10. The method according to claim 9, wherein the sub-stage of initial segmentation is performed basing on the analysis of the four sequential statistical feature vectors to define where the dividing marker has to be placed.

11. The method according to claim 10, wherein said analysis of the four sequential statistical feature vectors is performed in three steps, first of which may signalize about installation of the dividing marker, in this case other steps do not performed, the second of these steps may signalize about the absence of the dividing marker, in this case the third of said steps does not performed, and the third of said steps signalize about absence or installation of the dividing marker.

12. The method according to claim 11, wherein the first of said steps includes calculation of the modules of differences between two sequential coordinates of the first said sub-vector of the statistical feature vectors, comparison of the calculated values with the predefined threshold values, and the installation of the dividing marker if at least one said modulo is greater than the corresponding threshold value

13. The method according to claim 11, wherein the second of said steps includes calculation of the modules of differences between sum of two sequential coordinates of the second said sub-vector of the statistical feature vectors and sum of next two sequential coordinates of the second said sub-vector of the statistical feature vectors, comparison of the calculated values with the predefined threshold values, and signalizing about the absence of the dividing marker in the case when all of said modules are smaller than the corresponding threshold values.

14. The method according to claim 11, wherein the third of said steps includes calculation of the modules of differences between two sequential coordinates of the second said sub-vector of the statistical feature vectors, comparison of the calculated values with the predefined threshold values, the installation of the dividing marker if at least one said modulo is greater than the corresponding threshold value, and signalizing about the absence of the dividing marker in the opposite case.

15. The method according to claim 9, wherein the sub-stage of accurate segmentation is performed basing on the results of the initial segmentation and the analysis of sequence of said formant frequencies calculated for frames close to the dividing marker.

16. The method according to claim 15, wherein the analysis is based on calculation of the set of values of modules of differences between sums of mean values of the formant frequencies, wherein each modulo is calculated for the two intervals close to the dividing marker.

17. The method according to claim 16, wherein the intervals are sequential not-overlapped intervals, each of them has one and only one fixed border, and the border between said intervals is varied when the calculation of the set of values of the modules is performed.

18. The method according to claim 16, wherein the result of the accurate segmentation is the new position of the dividing marker, which corresponds to the maximum value of the values of modules of differences between sums of mean values of said formant frequencies.

19. The method according to claim 9, wherein goal of the sub-stage of the internal markers definition is a definition of two internal markers, which determine a most homogeneous interval inside each segment that was received at the sub-stage of accurate segmentation.

20. The method according to claim 19, wherein the two internal markers correspond to left and right halves of the segment and are calculated independently using calculation of the set of values of modules of differences between sums of mean values of the formant frequencies for each half of the segment.

21. The method according to claim 20, wherein the values of modules of differences are calculated for two not-overlapped intervals, each of them covers corresponding half of said segment, and the border between these two intervals is varied when the calculation of the set of values of the modules is performed.

22. The method according to claim 21, wherein each of the internal markers corresponds to the maximum of the values of modules of differences.